Ben Shapiro, Gavin Baker, Phil Deutch: OpenAI's GPT-5 Flop, AI's Unlimited Market, China's Big Advantage, Rise in Socialism, Housing Crisis

The episode opens with the host explaining that the show was meant to take a week off, but the independent moderator refused and announced a Friday recording with or without the regular co-hosts. In the banter that follows, someone complains that the co-hosts dropped out at short notice, with estimates ranging from five to ninety minutes before recording. David Friedberg is present, and Gavin Baker joins from the beach. The host introduces Baker as managing partner and CIO of a fund that invests in both public and private companies.

The host also introduces Ben Shapiro, host of the Ben Shapiro Show and co-founder of the Daily Wire, and mentions his book due in September, "Lions and Scavengers." On the show the book is described as being about civilization breaking down between "people who wanna build and people who wanna tear down". It is also said to be more personal than the author's usual work, having grown out of journaling, including after debates at Oxford University a month after October 7. The fifth guest, introduced by the host, is Phil Deutsch of NGP Energy Technology Partners. The host calls him the smartest person they know on energy, and he describes his family as government employees or public servants rather than politicians.

Before the main topics, the host promotes the All-In Summit and lists announced speakers: the CEOs of Eli Lilly, Uber, Arm and Zipline, along with Alibaba co-founder Joe Tsai, investors from Thoma Bravo and Sequoia, Cathie Wood, and science YouTuber Cleo Abram. The host adds that more names will be announced in the days before the event.

The host frames OpenAI's GPT-5 release, which came with two open-weight models, as "a bit underwhelming." The host describes the launch presentation as messy, with charts that were "flat and wrong" and had miscalculations and typos. In fairness, the host notes, the model topped some leaderboards, including the chatbot arena. The host contrasts the reaction with the earlier, bigger jumps of GPT-3.5 and GPT-4. Part of the disappointment, the host suggests, came from hype: Sam Altman had tweeted a Death Star image, which built expectations of something "otherworldly." (A tangent follows in which the panel ranks Star Wars films. One participant argues the franchise is built on an anti-technology, anti-progress worldview, with the Death Star as technology used for evil.)

Asked whether large language models are reaching diminishing returns, one panelist says that may be true "for OpenAI" but not for the field. They point to Grok 4 over Grok 3 as "a pretty giant leap." Their main point is historical: in the past, whenever OpenAI was overtaken, it answered with a clearly better model. They cite o3 arriving six to eight weeks after Grok 3 as an example. GPT-5 breaks that pattern.

this is the first time that, you know, OpenAI has released a new model that was not decisively the best.

The same speaker explains "benchmark saturation": models have become smart enough that tests designed even for human graduate students no longer measure their intelligence. Newer benchmarks were built in response. On Humanity's Last Exam, described as extremely hard questions across many disciplines, the speaker gives Grok 4 at 44.4% and GPT-5 at 42. On ARC-AGI-2, which targets the "jagged frontier" with tasks easy for humans but hard for AI, such as geometric puzzles, they say Grok 4 is significantly better than GPT-5. For balance, they note that GPT-5 is "very narrowly ahead" on the Artificial Analysis index. They also suggest that losing talent "may be beginning to weigh" on OpenAI.

The speaker discloses a possible bias. Their firm is a shareholder of xAI, and they add, "to be human is to be biased." They say they are optimistic about Grok 5, citing xAI having the largest Blackwell cluster, and that they expect an impressive new Gemini model. The reading of GPT-5's position therefore comes from an interested party and depends on competitors' releases that had not yet come out.

The host asks why OpenAI would ship a model that was not a clear winner, and whether that points to trouble innovating or a brain drain. A different view follows. The speaker says they do not know how much brain drain matters. For them, the key change is that GPT-5 picks the underlying model and reasoning path itself, where Pro users previously had to choose from a drop-down. They tested this with a complex research question and a simple information-retrieval question. The simple question was answered quickly, and the research question triggered thinking mode.

You don't want users going in and technically choosing which model.

In this view, the upgrade is mainly about user experience: "No one understands what those letters and numbers mean." The speaker adds that this helps explain why OpenAI is offering research functions more widely on a limited-use basis. A reply agrees but adds a caveat that changes how the early benchmarks should be read. The router that selects models was apparently broken, or did not work consistently for all users, for the first twelve to sixteen hours after release, so scores may improve.

The discussion ends this thread with OpenAI's strength in product and distribution. A participant cites Eric Fisher, a venture capitalist at Benchmark, who posted on X that OpenAI might still be a very valuable company even if it never released another model, because of its user base, product skill and distribution.

Asked about energy investing, one participant says the firm uses the models "all the time" for investing and finds them very efficient, just as any investor would. On the hardest energy problems and trade-offs, though, they place AI in "the first half of the first inning." As an example of work in that direction, they cite NVIDIA's climate model of the Earth.

Asked about the Daily Wire, another participant says AI is entering workflows through fast information gathering and editing tools. They mention a razor commercial released about two months earlier that was made entirely with AI. They say it still had an "AI slop feel to it a little bit" but was much cheaper than a live shoot.

The same participant, calling it the "dumb guy question," asks the central economic question: will productivity gains arrive fast enough to justify the money pouring in? They note that roughly 70% of S&P 500 gains are coming from the Magnificent Seven. They compare this with the internet, which they say was working by 1999 but took until roughly 2006 or 2007 to deliver its productivity gains, and ask whether "a crater" lies ahead.

The host sets up the answer with rough numbers. Low hundreds of billions of dollars a year are being spent now, and the host expects a trillion or two on data centers over the next five years. For the business market alone, the host suggests $50 to $100 per business user across a billion business users, and calls the result a 5 to $10 trillion TAM. The figures as spoken do not quite add up, since the intermediate revenue is stated as "a billion dollars." The host also asks whether there is enough capital in the world for investors to be patient.

The reply avoids giving a precise TAM, saying it is hard to estimate this early. It turns instead to measurable returns. Most of the big spenders are public companies that report quarterly, and their return on invested capital has gone up since the AI investments began. The speaker says early gains came mostly from operating-cost savings, "silicon OpEx to replace human OpEx." More recently they see revenue speeding up. They cite Meta's quarter, where AI kept users engaged longer and improved ad targeting so advertisers got better ROAS and spent more. They also mention Microsoft's Copilot adoption figures, while saying Copilot is "not the best product."

The speaker's main contrast with the dot-com era is about use. Much of the money in the so-called tech bubble was really a telecom bubble, spent laying "dark fiber" that went unused in the expectation that demand would come later. AI infrastructure, they argue, is being used to the limit.

the GPUs are used so intensively that they melt.

We're not building ahead of demand. In fact, we're clearly still behind demand,

The host adds that model training has no real equivalent in the dot-com era. The host also admits the comparison is loose, while noting that overbuilding did contribute to the dot-com bust.